[realmarcin]

The PHENIX 2.0-5936 conda installation on both M1 and M5 prints an empty-field banner when the inherited version variables are removed. Its own phenix_env.sh supplies the version. This adds a configured-install setup adapter so consistent vendor-derived identity can pass the refinement-cache measured-build gate while bare empty-field output remains unknown and non-reusable.

Trust contract:

• Clear inherited PHENIX_VERSION and PHENIX_RELEASE_TAG, plus shell startup hooks and exported functions, without changing the parent environment.
• Resolve the configured installation, require its regular phenix_env.sh inside the prefix, and reject missing setup, symlink escapes and PATH fallback. Source it in a child through fixed shell syntax with the path passed as a positional argument.
• Transfer only the two setup-derived version fields to that installation's resolved phenix.version. Accept one consistent Version/Release pair in one banner; reject contradictory, duplicate or mixed banners and fields split across output streams.
• Bind setup and dispatcher paths, SHA-256 digests, argv, output and parsed fields into evidence labeled vendor_setup_environment. The cache gate checks their structural and semantic consistency; directory hints and expected-version constants cannot authorize reuse.
• The evidence identifies the setup and dispatcher, not all installed libraries. It does not change scientific runner environments or reauthenticate historical results.

The change also isolates benchmark metadata unit tests from installed tool probes (#955) and aligns shell-function removal receipts with cache admission (#956). Review fixes validate directory symlink hops (#958), remove an unreliable shell-source assignment heuristic while preserving final-field/banner checks (#959), retain unknown evidence for path loops (#960), preserve timeout output (#961), restore executable availability semantics (the regression recorded on #409), and accurately document observation limits (#962).

Validation at 0af13c4037f94510306f65a4f8c385cec6fb1bf6: the required uv run --locked --extra benchmark -- bash scripts/validate.sh gate passed on Python 3.12. All 18 independent synthetic acceptance checks pass at runtime commit 679c949; the final documentation commit leaves those runtime files byte-identical. The controls cover cache reuse, inherited function-shaped values, directory symlink escapes, path loops, setup comments, banner contradictions, timeout evidence and executable availability. Fifteen cases failed on the preceding reviewed head, confirming regression sensitivity. The read-only M5 default resolver, with no PROTSTRUCT_* override or inherited version labels, reported 2.0-5936 from vendor_setup_environment and passed the cache measured-build gate. The bare control remained unknown/non-reusable. Setup and dispatcher SHA-256 values and the parent environment were unchanged. No refinement, benchmark, data download, paid provider call or vendor-install modification is part of this change.

Closes #954.
Closes #955.
Closes #956.
Closes #958.
Closes #959.
Closes #960.
Closes #961.
Closes #962.

Related: #760 still requires a second PHENIX build and an explicitly authorized scientific reproducibility run. The other existing scientific follow-ups remain unchanged. Merge requires separate user approval.

Remaining follow-up: #963 records pre-existing canonicalization of an OS-invalid regular-file/.. dispatcher alias. It reproduces on original main f3cfad3, resolves to an inside-prefix dispatcher, and does not affect the tested vendor layout. It remains open for a separate path-validity fix.

[realmarcin]

Fresh-context read-only review of f3cfad3..f37a017 completed. PRIMARY independently reproduced every actionable finding with temporary fake tools only. All are in scope for this adapter; no scientific decision or benchmark is required:

• P2: filtered legacy shell functions make valid vendor build evidence non-reusable #956: remaining producer/gate overlap for function-shaped inherited version variables; recorded on the existing issue.
• P2: PHENIX dispatcher resolution hides outside directory symlink hops #958: outside directory symlink hops are hidden when resolution returns inside the prefix; breaks the explicit containment contract.
• P2: PHENIX setup assignment scan misreads shell comments and quoted exports #959: source-text assignment scan misreads comments/quoted exports; remove the unreliable extra guard while retaining strict final setup/banner agreement and duplicate-banner rejection.
• P2: PHENIX build probe leaks parent symlink-loop exceptions #960: pre-existing parent symlink-loop exception escapes instead of producing unknown evidence; address while repairing the resolver.
• P3: PHENIX probe timeout receipts discard partial stdout and stderr #961: timeout receipts lose partial stdout/stderr; retain failure diagnostics without granting measured identity.
• PR #403: external tool availability is true for empty configured paths #409: regression of executable availability for mode-0644 dispatchers; recorded on the existing closed issue rather than duplicating its contract.

The review's seven focused suites passed, demonstrating these were coverage gaps. Fixes will use ordinary commits, followed by primary verification, the full QC gate and a fresh read-only review of the fix diff. PR remains unmerged.

[realmarcin]

Review fixes are committed locally as 679c949a67b2c90726ceab2464f823da959348b1. PRIMARY's independent synthetic matrix now passes all 18 cases (15 failed on the prior reviewed head); the secondary's focused suites and full locked gate pass.

One documentation-only finding remains: #962 corrects the audit's overstatement of permission/transient-change checks. The intended fix describes actual path/hash observations and the separate availability permission check. No new code guard or scientific run is required. Final push, primary QC and fresh fix-diff re-review follow that correction; this is not final acceptance or merge approval.

[realmarcin]

Final acceptance at 0af13c4037f94510306f65a4f8c385cec6fb1bf6:

• PRIMARY's exact-head locked Python 3.12 benchmark-extra QC passed; both validate (ubuntu-latest) and validate (macos-latest) are SUCCESS, and merge state is CLEAN.
• All 18 independent synthetic acceptance controls pass. Fresh full/fix/documentation reviews are complete with no unresolved in-scope findings.
• The M5 default version-only probe reports 2.0-5936 from vendor_setup_environment; the refinement-cache measured-build gate accepts it. Bare probing remains unknown. No overrides or inherited version labels were used, and vendor hashes are unchanged. The tested runtime is byte-identical at this head.
• Fixes cover P2: PHENIX version probe lacks build fields on both M1 and M5, preventing authenticated cache reuse #954, P2: benchmark environment unit test invokes installed tools during hermetic QC #955, P2: filtered legacy shell functions make valid vendor build evidence non-reusable #956, P2: PHENIX dispatcher resolution hides outside directory symlink hops #958, P2: PHENIX setup assignment scan misreads shell comments and quoted exports #959, P2: PHENIX build probe leaks parent symlink-loop exceptions #960, P3: PHENIX probe timeout receipts discard partial stdout and stderr #961, P3: PHENIX adapter audit overstates permission and transient-change checks #962 and the PR #403: external tool availability is true for empty configured paths #409 availability regression. The eight intended closing references are verified through the complete GraphQL connection.
• P3: PHENIX path canonicalization accepts file/../ dispatcher aliases #963 remains a separate P3 follow-up; the seven existing scientific follow-ups are unchanged. No refinement or benchmark ran.

The PR remains unmerged pending the user's explicit decision. The branch/worktree are retained clean and fully pushed, and the secondary is idle.